Event

RepresentNav Navigation Foundation Model Paper Released

1 reports · 1 sources

Jiawei Chen and other authors released a paper proposing the navigation foundation model RepresentNav: it uses Representation Residuals three-layer transformer to aggregate full-depth features from multiple visual foundation models, constructing a representation space driven by the policy's own objectives. The authors claim that on the EVT-Bench distracted tracking task, a single camera achieves 85.5% (+10.1); the "dream" in the model reduces collisions by a quarter, and occupancy was never supervised; the gains transfer to real robots with clutter and pedestrians.